LLM Hallucination Detection
Investigating factuality and attribution in large language models
Research on hallucination, attribution, and robustness of large language models including GPT-4, Gemini, and DeepSeek under distribution shift. Developing automatic and human-aligned evaluation frameworks for factuality, reasoning consistency, and failure analysis in scientific text generation.
Affiliation: Li Lab, Carnegie Mellon University
Advisor: Prof. Lei Li
Duration: Sep 2025 - Present
Research Focus
- Hallucination Detection: Identifying when models generate false or unsupported information
- Attribution Analysis: Evaluating how well models cite and support their claims
- Distribution Shift Robustness: Testing model reliability under varying input conditions
- Evaluation Frameworks: Building automated tools for assessing factuality and reasoning
Impact
This work aims to make LLMs more trustworthy and reliable for knowledge-intensive applications, particularly in scientific and technical domains where accuracy is critical.